Safety and Nonoptimal Usage of a Protected Intersection for Bicycling and Walking: A Before-and-After Case Study in Salt Lake City, Utah

: This paper describes a before-and-after case study of a protected intersection in Salt Lake City, Utah. The intersection was completed in late 2015 and represented one of the ﬁrst examples of a protected intersection design in North America. We analyzed bird’s-eye view video data that was recorded before the intersection was implemented and compared it against video data recorded from the exact same location after implementation. In order to examine changes in intersection usage and behavior, we operationalized safety in terms of the frequency of nonoptimal behaviors demonstrated by active transportation modes. We found that active transportation usage of the intersection has increased since the new conﬁguration, with most of that growth attributable to e-scooter users. There was minimal change in the rates of nonoptimal behaviors by pedestrians. Bicyclists showed mostly decreased rates of nonoptimal behaviors, suggesting improved safety for this mode. E-scooter users, however, demonstrated nonoptimal behaviors at very high rates as compared with other active modes. This case study gives evidence that a protected intersection can have positive e ﬀ ects on active transportation volume and safety in a U.S. context.


Introduction
As a sustainable urban transport mode and important type of physical activity, bicycling has been recognized to provide various health benefits, including a reduction in the risks of diabetes, heart disease, obesity, cancers, and an increase in metabolic and cardiorespiratory functions [1][2][3][4]. In the past few decades in the USA, bicycle trips have increased significantly through the addition of new infrastructure and a general increase in attention given to active transportation [5,6]. However, as increased bicycle usage involves increased conflicts between bicyclists, pedestrians, and motor vehicles, creating safe environments for active transportation and physical activity has been a critical mission of planners. Whether streets incorporate bicycle lanes is one of the most significant environmental factors that promotes cycling behaviors and decreases bicycle crashes [3,7]. As governments develop a wide variety of bicycle-friendly infrastructures and programs [5], an important research question is whether, and to what extent, each intervention impacts bicycling and road safety.
Protected intersections are one of the novel infrastructure designs that incorporates many of the following attributes of successful European designs: physically separating rights of way, making pedestrians and bicyclists more visible to drivers, providing shorter crossing distances for active transportation users, and prioritizing active modes through signalization. These intersection treatments are relatively new to North American cities, and there are only a handful of protected intersections in North America [8]. While there are differences in design features among the existing protected intersections, Salt Lake City's design is understood to be the best example of an intersection that incorporates the most aspects of the adapted North American typology [8]. However, to date, no studies have empirically investigated the impact of protected intersections on cycling and other users' behaviors in the U.S. context. This study utilizes before-and-after video analysis to examine the effects of the implementation of a protected intersection on safety and usage by pedestrians, cyclists, and other active users. The study site is the signalized intersection at 200 West 300 South of Salt Lake City in Utah, which was completed in its current form in 2016. With video data of the intersection taken in 2015, 2016, and 2018, we focus on two outcomes, i.e., changes in intersection usage and changes in nonoptimal behavior. First, we identified how many people used the intersection and how the figures have changed since the new intersection was introduced. Second, we investigated the proportion of people who did not comply with safe utilization of the intersection and examined whether the nonoptimal behavior rates have changed since the new intersection was introduced.
This study is the first of its kind for a U.S. protected intersection, and its insights should help to guide active transportation infrastructure design and public-health practices in the future. Leading researchers in bicycle and pedestrian planning have written at length about the lessons that can be learned from European traffic designs, and practitioners have adapted those designs to conform to an American context. We know a fair amount about the effects of these traffic configurations from studies that have examined them empirically, however, it cannot necessarily be assumed that the findings are generalizable to the North American context. Our in-depth case study on a protected intersection should bring critical insights that help researchers and practitioners to better understand the effects of intersection design changes in the U.S. context.

Literature Review
When active transportation planning first came to the fore in North America, planners looked to European cities for insight into how to create more hospitable environments for pedestrians and bicyclists. Policies that can help make active transportation more appealing than automobiles often include gasoline or carbon taxes, high registration fees for vehicles, and other economic disincentives [9,10]. Such policies, however, might be considered to be untenable within the context of American politics. If the stick is not a viable option for U.S. planners and policymakers, what about the carrot? Pucher and Buehler [11] looked to European examples and found that, beyond automobile restricting policies, countries such as the Netherlands, Denmark, and Germany relied on ubiquitous provision of safe infrastructure for bicyclists and pedestrians in order to enhance the attractiveness of these modes. European cities have long provided for safe integration of bicyclists and vehicles through the use of special signalization, physical separation, advanced stop location for bicyclists, and other designs that facilitated safe use of streets among multiple modes of transportation [12]. Researchers have found that physical and visual separation of automobile and bicycle rights of way improved the perceived safety of active transportation and have led to increased usage [3,11,[13][14][15]. Additionally, the benefits of investment in active transportation infrastructure have been shown to far exceed the costs. Gotschi [16] determined that an investment of bicycles of USD 605 million, in Portland, OR, would result in health care savings of USD 594 million, fuel savings of USD 218 million, and value of statistical lives of USD 12 billion over a 50 year period.
Through the addition of new infrastructure and a general increase in attention given to active transportation in the past few decades in the USA, bicycle trips have more than doubled [5,6]. While this represents a terrific realization of the efforts of planners to enhance the environment for active transportation, it also presents a new opportunity for increased conflicts between bicyclists, pedestrians, and motor vehicles. Motor vehicles, on their own, are immensely dangerous, with motor vehicle crashes being a leading cause of death globally, and among the top ten most prominent causes of death among young people nationally [17,18]. The best estimates of pedestrian-vehicle crashes suggest that there are more than 100,000 of these types of incidents per year in the USA, with nearly 5000 of them resulting in a fatality [19,20]. Although there is always potential for conflict among automobiles, bicyclists and pedestrians, the likelihood of a collision is far greater at intersections, where all modes interact using much of the same space [21][22][23]. In fact, a study of 15 years of crash data in Palo Alto, CA, found that nearly three-quarters of crashes involving cyclists and motor vehicles occurred at intersections [22]. For this reason, much of the literature on safe design for bicycles and vehicles has focused on behavior at intersections [24][25][26][27][28]. There has also been a tendency to study crash history, as longitudinal data have some advantages in their ability to imply causation. Unfortunately, however, these studies are limited because their design necessitates decades worth of data [28][29][30].
There are many design features that have been developed in hopes of creating a safer environment for the interaction of pedestrians, bicyclists, and vehicles. Colored markings are one of the simplest and most cost-effective methods used for delineating the respective spaces for bicycles and automobiles [31]. Colored markings through intersections demand even more attention from roadway users and have been shown to decrease intermodal crash rates in both European and American contexts [32,33]. Bicycle boxes, or protected areas where bicycles safely enter and stop in an advanced position in an intersection in front of automobiles, are intended to make bicyclists more visible to drivers. However, limited research has shown that this design provides scant evidence for increased safety. Hunter [33] found little reduction in conflicts between automobiles and bicycles and a significant amount of encroachment into the bicycle box by vehicles. New methods for traffic signalization have also been employed in an attempt to improve safety at intersections. Scramble signals allow bicycles and pedestrians to move freely in all directions while automobiles are restricted from entering the intersection. Bicycle-only directional signal phasing also restricts automobiles from entering the intersection while bicycles and pedestrians are permitted, but they also limit the movements of active users for increased safety. The methods have both been shown to improve the safety of intersections by reducing the number of conflicts between bicyclists and vehicles [21,34]. Increases in intersection usage by active transportation modes have also been attributed to scramble signals [34]. Finally, as these methods make intersections more appealing to active transportation users, simply the increased number of pedestrians and bicyclists using the intersection makes the environment safer for those modes. Jacobsen [35] found that motorists were less likely to collide with pedestrians and bicyclists when more of these users were present.
Although the above findings help planners and traffic engineers determine what treatments might improve safe integration of all modes of transportation at intersections, there remains a significant amount of knowledge yet to be discovered. Many European cities such as Copenhagen and Amsterdam have developed intricate systems of protected bicycle lanes and intersections that employed many of the methods we have discussed for creating safer intersections. However, studies of these intersections are limited in number, and confounding the situation further is the fact that findings from studies of European designs in European contexts might not be applicable in a North American context. There is a dire need, therefore, for studies that examine the effects of protected intersection treatments in the USA. This study examines the best American example of a protected intersection to date, using before-and-after video analysis of pedestrian, bicycle, vehicle, and other users in the intersection.

Research Design
This study employs a before-and-after case study, also called quasi-experimental design or natural experiment. While there have been studies of foreign examples of new intersection designs [36,37], the findings from these studies might not necessarily be applicable to an American context [21]. Salt Lake City's 200 West 300 South intersection is the best example of the modern protected intersection design among the twelve that currently exist in the USA. Thus, a longitudinal analysis of this example should help researchers and practitioners better understand the effects of intersection design changes in the U.S. context.

Operationalizing Safety and Tested Hypotheses
Safety is a relatively nebulous construct that has been measured in countless ways throughout the vast transportation planning and engineering literature. A common way to measure safety is the incidence or frequency of collisions. This measure is effective when considering the safety of a traffic corridor, traffic analysis zone, or larger geography over time. However, when we refine the spatial or temporal scale to a single intersection at a given moment, crash rates become too infrequent to provide meaningful analysis. The availability of unique high-quality video data taken over three different points in time provides us an opportunity to measure safety in a completely novel way with greater detail and nuance.
For the purposes of this study, we are operationalizing safety as the prevalence of what we call "nonoptimal behaviors". Optimal behaviors are considered to be behaviors by people who conform to the expected use of given lanes at the protected intersection, and thus prevent possible conflicts ( Figure 1). Thus, it can be assumed that if road users are changing their behavior to a higher degree of conformity than the expected optimal behaviors, the examined area will be a safer place for road users. Thus, in this study, nonoptimal behaviors are defined as any behaviors that are deviant to the optimal behaviors. This is different from illegal traffic maneuvers because not all dangerous intersection movements are necessarily illegal. Protected bicycle intersections, for example, are designed to reduce the amount of space and time that pedestrians and cyclists are exposed to vehicular traffic. This design, in part, is intended to facilitate safer left turn movements of cyclists by allowing them to cross the intersection straight ahead, wait for the next signal phase in a protected bike box, and then cross again straight ahead. Making a left turn in this way greatly reduces the risk to cyclists of colliding with a vehicle by reducing their exposure to oncoming vehicles. However, making an exposed left turn is not illegal for cyclists. This is the best example of the difference between an illegal maneuver and a nonoptimal behavior.
Sustainability 2020, 12, x FOR PEER REVIEW 4 of 14 this example should help researchers and practitioners better understand the effects of intersection design changes in the U.S. context.

Operationalizing Safety and Tested Hypotheses
Safety is a relatively nebulous construct that has been measured in countless ways throughout the vast transportation planning and engineering literature. A common way to measure safety is the incidence or frequency of collisions. This measure is effective when considering the safety of a traffic corridor, traffic analysis zone, or larger geography over time. However, when we refine the spatial or temporal scale to a single intersection at a given moment, crash rates become too infrequent to provide meaningful analysis. The availability of unique high-quality video data taken over three different points in time provides us an opportunity to measure safety in a completely novel way with greater detail and nuance.
For the purposes of this study, we are operationalizing safety as the prevalence of what we call "nonoptimal behaviors". Optimal behaviors are considered to be behaviors by people who conform to the expected use of given lanes at the protected intersection, and thus prevent possible conflicts ( Figure 1). Thus, it can be assumed that if road users are changing their behavior to a higher degree of conformity than the expected optimal behaviors, the examined area will be a safer place for road users. Thus, in this study, nonoptimal behaviors are defined as any behaviors that are deviant to the optimal behaviors. This is different from illegal traffic maneuvers because not all dangerous intersection movements are necessarily illegal. Protected bicycle intersections, for example, are designed to reduce the amount of space and time that pedestrians and cyclists are exposed to vehicular traffic. This design, in part, is intended to facilitate safer left turn movements of cyclists by allowing them to cross the intersection straight ahead, wait for the next signal phase in a protected bike box, and then cross again straight ahead. Making a left turn in this way greatly reduces the risk to cyclists of colliding with a vehicle by reducing their exposure to oncoming vehicles. However, making an exposed left turn is not illegal for cyclists. This is the best example of the difference between an illegal maneuver and a nonoptimal behavior. We contend that the frequency of nonoptimal behaviors is an effective proxy for safety, especially when studying the interaction of multiple transportation modes at a protected intersection. Over the three-year study period, there were only four reported traffic collisions near the intersection. The infrequency of this event and the lack of detail related to the cause of the collision gives us little information about how the configuration of the intersection is leading to safety for active transportation users. For this reason, we propose nonoptimal behavior as an effective measure of safety for analyzing a protected intersection.
In this study, we test two hypotheses. First, the implementation of a protected intersection promotes active transportation usage. The protected intersection design physically separates cyclists and pedestrians from cars, where we expect active transport users to feel more comfortable using the intersection area, and thus their usage will increase. Second, the implementation of a protected intersection reduces the rates of nonoptimal behaviors of active transport users. The protected We contend that the frequency of nonoptimal behaviors is an effective proxy for safety, especially when studying the interaction of multiple transportation modes at a protected intersection. Over the three-year study period, there were only four reported traffic collisions near the intersection. The infrequency of this event and the lack of detail related to the cause of the collision gives us little information about how the configuration of the intersection is leading to safety for active transportation users. For this reason, we propose nonoptimal behavior as an effective measure of safety for analyzing a protected intersection.
In this study, we test two hypotheses. First, the implementation of a protected intersection promotes active transportation usage. The protected intersection design physically separates cyclists and pedestrians from cars, where we expect active transport users to feel more comfortable using the intersection area, and thus their usage will increase. Second, the implementation of a protected intersection reduces the rates of nonoptimal behaviors of active transport users. The protected intersection provides more exclusive places for individual types of active transport users, in which we expect fewer conflicts and fewer nonoptimal behaviors.

Data Collection and Processing
We obtained video records of intersection usage from the Salt Lake City Transportation Division. Salt Lake City had collected a total of 36 h of video data. These data came from 12-h segments taken on the following three separate days: 31 July 2015 (before), 19 August 2016 (after 1), and 24 August 2018 (after 2) ( Figure 2). Each video was recorded under particular controlled conditions, i.e., taken for 12 h (7:00 to 19:00) at the same place, on the same day of the week (Friday), in the same time of the year (July or August), in similar weather conditions. Sustainability 2020, 12, x FOR PEER REVIEW 5 of 14 intersection provides more exclusive places for individual types of active transport users, in which we expect fewer conflicts and fewer nonoptimal behaviors.

Data Collection and Processing
We obtained video records of intersection usage from the Salt Lake City Transportation Division. Salt Lake City had collected a total of 36 h of video data. These data came from 12-h segments taken on the following three separate days: 31 July 2015 (before), 19 August 2016 (after 1), and 24 August 2018 (after 2) ( Figure 2). Each video was recorded under particular controlled conditions, i.e., taken for 12 h (7:00 to 19:00) at the same place, on the same day of the week (Friday), in the same time of the year (July or August), in similar weather conditions. We divided the video analysis into two separate tasks (see Figure 3). First, we counted intersection users, segmenting counts by mode of travel. We ran the video at a faster than real-time speed to expedite the analysis for this task. We analyzed all twelve hours of video for each day, and recorded counts by the hour. Second, we counted nonoptimal behaviors and calculated the nonoptimal behavior rates to the total counts. This required running the video at near real-time playback speeds with frequent pausing and rewinding to ensure that behaviors were identified correctly, and nothing was missed. Because this process was so intensive, we sampled three one-hour segments (09:00-10:00, 13:00-14:00, 17:00-18:00) for each observation day. We chose these times specifically because they were the peak periods for active transportation usage, based on our preliminary analysis of activity in the intersection.
We also included vehicle traffic data in our analysis, however, measured at a courser scale of geography. We included annual average daily traffic (AADT) for one of the roads in the intersection, as well as crash data provided by the Utah Department of Transportation. These data provided useful insight into how the new intersection configuration had affected vehicle traffic and safety. This is useful context as we discuss active transportation usage and safety. We divided the video analysis into two separate tasks (see Figure 3). First, we counted intersection users, segmenting counts by mode of travel. We ran the video at a faster than real-time speed to expedite the analysis for this task. We analyzed all twelve hours of video for each day, and recorded counts by the hour. Second, we counted nonoptimal behaviors and calculated the nonoptimal behavior rates to the total counts. This required running the video at near real-time playback speeds with frequent pausing and rewinding to ensure that behaviors were identified correctly, and nothing was missed. Because this process was so intensive, we sampled three one-hour segments (09:00-10:00, 13:00-14:00, 17:00-18:00) for each observation day. We chose these times specifically because they were the peak periods for active transportation usage, based on our preliminary analysis of activity in the intersection.
We also included vehicle traffic data in our analysis, however, measured at a courser scale of geography. We included annual average daily traffic (AADT) for one of the roads in the intersection, as well as crash data provided by the Utah Department of Transportation. These data provided useful insight into how the new intersection configuration had affected vehicle traffic and safety. This is useful context as we discuss active transportation usage and safety. Sustainability 2020, 12, x FOR PEER REVIEW 6 of 14

Types of Nonoptimal Behaviors
Although counting the number of intersection users is quite straightforward, as noted in the previous section, identifying and counting nonoptimal behaviors is challenging. Thus, to increase the precision of the work, a pilot study was conducted, in which we identified a total of 19 nonoptimal behaviors at the intersection, seven for bicyclists, five for pedestrians, and seven for others (Table 1). Examples included walking on bikeways, riding a bike in the opposite direction on bikeways, and disobeying signals. First, as each user approached the intersection, we examined the number of users moving outside the correct lanes. Second, as each user turned or crossed the intersection, we counted the number of users crossing outside the correct crossing lanes, disobeying signals, or riding a bike in the opposite direction. Lastly, as each user stopped at the intersection, we investigated the number of users standing in wrong places. Figure 4 shows screenshots of examples of nonoptimal behaviors. Our assumptions, with respect to what constituted nonoptimal behavior, were based on our expertise in traffic safety and local and national policy. Additionally, our decisions for categorizing nonoptimal behaviors were made with the input of an external expert panel consisting of practicing planners and engineers.

Types of Nonoptimal Behaviors
Although counting the number of intersection users is quite straightforward, as noted in the previous section, identifying and counting nonoptimal behaviors is challenging. Thus, to increase the precision of the work, a pilot study was conducted, in which we identified a total of 19 nonoptimal behaviors at the intersection, seven for bicyclists, five for pedestrians, and seven for others (Table 1). Examples included walking on bikeways, riding a bike in the opposite direction on bikeways, and disobeying signals. First, as each user approached the intersection, we examined the number of users moving outside the correct lanes. Second, as each user turned or crossed the intersection, we counted the number of users crossing outside the correct crossing lanes, disobeying signals, or riding a bike in the opposite direction. Lastly, as each user stopped at the intersection, we investigated the number of users standing in wrong places. Figure 4 shows screenshots of examples of nonoptimal behaviors. Our assumptions, with respect to what constituted nonoptimal behavior, were based on our expertise in traffic safety and local and national policy. Additionally, our decisions for categorizing nonoptimal behaviors were made with the input of an external expert panel consisting of practicing planners and engineers.

Reliability of Observation
We tested inter-rater reliability, also called inter-observer reliability or inter-rater agreement, for the following two types of data: (1) total counts by transportation mode and (2) counts by each nonoptimal behavior type. The intraclass correlation coefficient (I.C.C.) is an appropriate measure of inter-rater reliability and has been used in behavioral research [38][39][40]. The I.C.C. measures the extent of agreement between two or more raters for continuous variables. Higher I.C.C. values indicate greater reliability, with an I.C.C. estimate of 1 indicating perfect agreement, 0 indicating only random agreement, and a negative I.C.C. indicating systematic disagreement.
According to Cicchetti [41], inter-rater reliability should be considered poor for I.C.C. values less than 0.40, moderate for values between 0.40 and 0.59, good for values between 0.60 and 0.74, and excellent for values greater than 0.75. Following the guideline from Koo and Li [39], we used a "twoway mixed model" because raters were fixed, and our subjects were chosen randomly and measured "absolute agreement" rather than consistency. The I.C.C. values were computed in R 3.6.0 software (psych package).
A fifteen-minute video segment was used as a unit of analysis for the reliability test. The minimum sample size for three raters with ICC 0.5 minimum, power 80%, and alpha 0.05 is 11 [42]. Thus, three observers watched the same 11 fifteen-minute video segments and counted bicycles, pedestrians, and other types of modes (e.g., scooter, skateboarders, segways, etc.) for each nonoptimal behavior type.

Reliability of Observation
We tested inter-rater reliability, also called inter-observer reliability or inter-rater agreement, for the following two types of data: (1) total counts by transportation mode and (2) counts by each nonoptimal behavior type. The intraclass correlation coefficient (I.C.C.) is an appropriate measure of inter-rater reliability and has been used in behavioral research [38][39][40]. The I.C.C. measures the extent of agreement between two or more raters for continuous variables. Higher I.C.C. values indicate greater reliability, with an I.C.C. estimate of 1 indicating perfect agreement, 0 indicating only random agreement, and a negative I.C.C. indicating systematic disagreement.
According to Cicchetti [41], inter-rater reliability should be considered poor for I.C.C. values less than 0.40, moderate for values between 0.40 and 0.59, good for values between 0.60 and 0.74, and excellent for values greater than 0.75. Following the guideline from Koo and Li [39], we used a "two-way mixed model" because raters were fixed, and our subjects were chosen randomly and measured "absolute agreement" rather than consistency. The I.C.C. values were computed in R 3.6.0 software (psych package).
A fifteen-minute video segment was used as a unit of analysis for the reliability test. The minimum sample size for three raters with ICC 0.5 minimum, power 80%, and alpha 0.05 is 11 [42]. Thus, three observers watched the same 11 fifteen-minute video segments and counted bicycles, pedestrians, and other types of modes (e.g., scooter, skateboarders, segways, etc.) for each nonoptimal behavior type. Table 2 shows that the inter-rater reliability of total counts is excellent for all modes of transportation in the observed intersection. Table 3 also illustrates that the reliability of nonoptimal behaviors is considered to be excellent by mode, with I.C.C. values from 0.84 for pedestrian to 0.94 for other users (scooter users, etc.), but there are greater differences in reliability by behavior type. When coded by each behavior type, 18 of the 19 categories had one or more incidences in the sample video clips, one category, "other modes crossing in street", was not observed by any rater, as was rarely found in our observation result. Out of 19 categories, 10 categories were considered to be in an excellent reliability range (ICC > 0.75), three in a good reliability range (0.60 < ICC < 0.75), and four in a fair range (0.40 < ICC < 0.60). One category that reported a poor I.C.C. value (less than 0.40) was "bicycle stopping in wrong place" (I.C.C. = 0.32). Thus, data for these low-reliability categories would need careful interpretation.

Statistical Analysis
To identify changes in behavior from before the protected intersection implementation and after, we employed the difference of means tests and descriptive statistics. The application of inferential statistics to our data depends on the sample size and number of observations that we have available. On the basis of these constraints, we only applied t-tests to intersection usage data. For nonoptimal behavior measurements, we used descriptive statistics. This meant that the observations of changes in nonoptimal behavior were simply descriptive, and we could not say that the intersection implementation led to statistically significant changes. Table 4 displays changes to intersection usage as estimated by video data counts of different modes. We calculate change as the difference between the average of the two post-implementation sample periods (2016 and 2018) from the pre-implementation period (2015). The exception to this method is the calculation of the change in the "other (scooter only)" category, which is measured as the difference between the average of the two days of video (August 2015 and 2016) before the deployment of shareable scooters in the summer of 2018 and the one post-deployment, as well (August 2018). Table 4. Intersection usage.

Mode
Year In total, we found an increase of 157 additional non-motorized users of the intersection after the implementation of the new protected intersection configuration, per the two days of video data analyzed. This is a modest increase, with the greatest growth in usage happening between 2016 and 2018. Interestingly, we see little change in bicycling and walking usage from before and after the implementation of the protected intersection. Most notably, almost all the growth in active transportation usage of the protected intersection between 2016 and 2018 is attributable to an increase in e-scooters. This mode demonstrates 336 more scooters between 2015-2016 and 2018 (p < 0.01). Table 5 displays measures of nonoptimal behavior among all active transportation modes. To calculate the rate of nonoptimal behaviors, we divided the counts of behaviors by the estimated usage counts from the previous analysis. We recorded the travel mode volume estimates by the hour, so the volumes were aggregated from the 09:00-10:00, 13:00-14:00, and 17:00-18:00 segments, corresponding to our nonoptimal behavior sample times.

Nonoptimal Behavior
Some patterns can be observed in nonoptimal behaviors among bicyclists after the implementation of the protected intersection. Bicyclists crossing the intersection in the crosswalk declined from a pre-implementation rate of 17.1% to a post-implementation average of 6.2%. This is likely due to the improved painted pavement that makes the area where bicyclists are expected to cross the intersection much clearer than in the previous configuration. Bicyclists also demonstrated a decrease in the rates of exposed left turns after the implementation of the protected intersection. This came as somewhat of a surprise, because we expected the optimal movement for a left turn, which required that a bicyclist travel straight across the intersection, then wait for another signal change, and then travel straight across again, to be too onerous to elicit an observable change even with the new configuration. This was not the case, however, and the rate of exposed left turns declined from 17.1% to an average of 2.5%. Note: The data from three one-hour segments (09:00-10:00, 13:00-14:00, and 17:00-18:00), in each of the three daily, August counts. All numbers were reported by one observer who used a reliable observation protocol. We did not report nonoptimal behaviors rates for the "other" category prior to the 2018 observation because the small number of observations contributed to misleading measures. However, the value of 2018 "other" data was found to make a comparison with nonoptimal bicycling behaviors. An asterisk mark (*) indicates nonoptimal behaviors with a poor I.C.C. value.
Bicyclists also demonstrated lower rates of stopping in incorrect positions after the implementation of the protected intersection. We attribute this improvement to the clearer delineation of space for stopped cyclists in an area known as the "bike box." This area is painted a different color than the other pavement and is also physically separated from the street, likely leading to an increased feeling of safety for bicyclists waiting for the changing traffic signal (Figure 1). There was one nonoptimal behavior that increased significantly after the protected intersection was implemented, i.e., bicyclists not obeying the signal and crossing against the movement of cars. This presents potential safety concerns as bicyclists are vulnerable to being hit from the side by a car traveling perpendicular to the direction of their travel. We suggest that this change is likely due to the reduction in travel lanes, and thus the shorter distance the bicyclist must traverse across the intersection while being exposed to motorized traffic.
Pedestrian behavior did not change in a way that provided many observable patterns. The change in nonoptimal behavior of this mode was small and did not seem to fluctuate in a way that reflected the impact of the protected intersection on pedestrian movements. One exception was the decrease in the rate of crossing outside crosswalks (6.1% to 3.9%) even though the crosswalks became narrower after the protected bike intersection was installed. We hypothesize that a clear delineation between a pedestrian crosswalk and a bike lane may encourage both pedestrians and cyclists to stay in their right-of-way.
As we discuss above in reporting changes in usage of the intersection over the study period, the arrival of e-scooters in Salt Lake City, between 2016 and 2018, led to changes in the way the protected intersection was utilized. We did not report figures for the "other" category prior to the 2018 observation, because the small number of observations contributed to misleading measures. For this reason, we simply compared 2018 observations of e-scooter nonoptimal behavior to 2018 bicyclist behavior. We made this comparison because existing regulations for e-scooter travel typically placed similar restrictions on this mode as were placed on bicycles. We see that e-scooters display higher rates of nonoptimal behavior in every category with the exception of riding in the roadway. E-scooters also demonstrate similar, but slightly higher rates of making exposed left turns and stopping out of place. E-scooter users were much more likely to disobey the signal as compared with their bicycling counterparts, with 16.8% of users crossing against cars' movement. The nonoptimal behavior that e-scooter users were most likely to exhibit was riding on sidewalks. A total of 43.2% of all observed e-scooter users were riding on sidewalks instead of the protected bicycle lane where they were expected to remain. Similarly, e-scooter riders also crossed the intersection within the crosswalk instead of crossing in the bicycle lane at a rate of 22.1% as compared with bicyclists' rate of 5.1%. Generally, e-scooter users were much more likely to exhibit nonoptimal behavior in their approach and interaction with the protected intersection.

Vehicle Traffic and Safety
Although we did not report changes in drivers' nonoptimal behaviors, there were no collisions or risky signal violations of drivers found during the observation period. Thus, we further examined whether there were differences in traffic volume and the number of collisions since 2016 (post-implementation) ( Table 6). The annual traffic data show that there was an increase of 1000 vehicles along the north-south roads at the intersection. This increase was similar to the mean increase in traffic volume of the streets that intersect Salt Lake City, implying there was no significant traffic volume change attributable to the new intersection design. Furthermore, there was no consistent increase in the number of collisions at the intersection during the examined post-implementation period. Four crashes occurred in the 2018 record only, and none of them involved pedestrians or cyclists, or any injury. A possible explanation is that the streets at the intersection became narrower to incorporate the protected intersection design elements (e.g., protected bike lanes, on-street parking, and center islands), possibly resulting in a reduction in traffic speed and no severe injury or fatal crashes. Nevertheless, the increase in the number of crashes is not consistent since 2016 and is not sufficient evidence to prove that the crash events were affected by the new intersection design.

Conclusions
This paper highlights examples of changes to usage of a protected intersection in Salt Lake City, UT. We found that after the implementation of a new protected intersection configuration, active transportation usage increased during our three-year study period from 2015 to 2018. Increases in active transportation usage during this time, however, were mostly attributable to a rapid spike in e-scooter users. There has been some speculation among transportation planners and academics whether this new mode might replace trips that would otherwise have been made by walking or bicycling. Initial findings from this limited case study suggest that if e-scooters are, in fact, cannibalizing active transportation trips, the transfer of trips to e-scooters is not significantly diminishing the number of bicycle and pedestrian users at this protected intersection.
We also analyzed the rates of nonoptimal behaviors at the protected intersection, determining how the frequency of these behaviors changed with the implementation of the protected intersection. We found that pedestrian behavior showed a slight change in response to the new configuration, i.e., higher rates of pedestrians were observed staying within the confines of the crosswalk while crossing. More noticeably, bicycle behavior responded to the new infrastructure with reductions in bicyclists crossing within the pedestrian crosswalk, stopping on sidewalks and in the street, and making exposed left turns. The reduction of these nonoptimal behaviors suggests a positive effect on safety for bicycles with the implementation of the protected intersection. Conversely, however, bicyclists tended to cross against the signal at higher rates with the new configuration.
A new user in the protected intersection space was the e-scooter rider. These riders demonstrated higher rates of nonoptimal behaviors than both pedestrians and bicyclists. They were more likely to perform all nonoptimal behaviors than their counterparts on two feet or bicycles except for making exposed left turns. E-scooters utilized the sidewalk at an exceptionally high rate as compared with bicyclists, with 43% of all users preferring this space. While we categorized this as a nonoptimal behavior due to guidance from regulations emerging around the country, this preference should be considered by planners when deciding how to deal with the proliferation of e-scooters. There may be reasons why e-scooter users prefer to ride on the sidewalk as opposed to the bicycle lane. Further research into the reasons for this behavior might elucidate the best ways to plan for and regulate e-scooter use in the future.
Generally, this study showed that active transportation use increased after the implementation of a protected intersection in Salt Lake City, UT. We also showed that many nonoptimal behaviors have been reduced since the new configuration was deployed. This case study gives some evidence that a protected intersection can have positive effects on active transportation volume and safety in a U.S. context. However, we must note that this case study does not assert causality related to the observed changes in volumes and behaviors. Although the before-and-after nature of our samples might be a type of a quasi-experimental design, the small number of samples, i.e., three observation days, limits the internal validity and external validity of our findings. More data and analysis are necessary to begin to make more concrete assertions about the relationships among active transportation volumes, behaviors related to public health and safety, and protected intersection configurations.
Additionally, the context of the intersection is very specific, which limits the ability to generalize beyond the treatment and its surrounding environment. In order to be able to generalize more broadly, future research needs to include more intersections in a wider variety of contexts. We suggest that future research could employ artificial intelligence techniques such as machine learning and computer vision algorithms to facilitate the analysis of greater volumes of data. Such techniques could potentially allow for the study of more intersections and more periods in time to increase the generalizability of research findings. Recent developments in urban transportation, such as the emergence of the COVID-19 pandemic and its effects on travel behavior, have increased the need to incorporate new data into the analysis of active transportation safety. Along with improved methods of observation and analysis, we suggest researchers update this study with more contemporary data.